Research map / Large language models / Instruction and fine-tuning
Differences That Matter: Auditing Models for Capability Gap Discovery and Rectification
Qihao Liu, Chengzhi Mao, Yaojie Liu, Alan Yuille, Wen-Sheng Chu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add AuditDM-generated data to fine-tuning | PaliGemma2-3B · AI2D | accuracy: 76.0 -> 85.3 | Improved |
| add AuditDM-generated data to fine-tuning | PaliGemma2-3B · ChartQA | accuracy: 54.0 -> 63.8 | Improved |
| add AuditDM-generated data to fine-tuning | PaliGemma2-3B · RefCOCO | accuracy: 76.3 -> 77.8 | Improved |
| add AuditDM-generated data to fine-tuning | Gemma3-4B · MMBench-v1.1 | accuracy: 67.6 -> 75.0 | Improved |
| add AuditDM-generated data to fine-tuning | Gemma3-4B · POPE | accuracy: 85.1 -> 85.5 | No clear effect |
| add fine-tuned auditor instead of prompt engineering for failure search | AuditDM · Failure search (20K attempts) | search success rate: 21.4% -> 91.1% | Improved |
| add probing-question component (ablation) | AuditDM PaliGemma2-3B 224px · GQA | accuracy: 66.2 -> 68.5 | Improved |
| add image generation component (ablation) | AuditDM PaliGemma2-3B 224px · GQA | accuracy: 66.2 -> 66.9 | Improved |
| add image editing component (ablation) | AuditDM PaliGemma2-3B 224px · GQA | accuracy: 66.2 -> 67.2 | Improved |
| combine the best auditing components (ablation) | AuditDM PaliGemma2-3B 224px · AI2D | accuracy: 74.7 -> 79.4 | Improved |
| add image generation or probing data on RefCOCO (ablation) | AuditDM PaliGemma2-3B 224px · RefCOCO | accuracy | Hurt |
| add auditor-targeted samples instead of fine-tuning on same images only (ablation) | Gemma3-4B · MMBench-v1.1 | accuracy: 69.6 -> 75.0 | Improved |
Closest papers by meaning
- Adaptive auditing of AI systems with anytime-valid guarantees (NeurIPS 2026 · similarity 0.84)
- AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems (NeurIPS 2026 · similarity 0.84)
- An Auditing Test to Detect Behavioral Shift in Language Models (ICLR 2025 · similarity 0.84)
- RvLLM: LLM Runtime Verification with Domain Knowledge (NeurIPS 2025 · similarity 0.84)
- Improving Audit Realism with Inference-Time Compute and Deployment Scaffolds (NeurIPS 2026 · similarity 0.84)
- e3: Learning to Explore Enables Extrapolation of Test-Time Compute for LLMs (ICLR 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
